Product layer / Building

The evidence created during translation.

The Synset Evidence Record preserves the model versions, gaps, interventions, protocols, results, limitations, and regression cases created during translation and hardening.

This is an illustrative product preview, not customer evidence, a regulatory record, or a generally available evidence-operations platform.

Primary schema

Follow the decision, not just the file.

  1. 01Objective
  2. 02Gap
  3. 03Intervention
  4. 04Protocol
  5. 05Result
  6. 06Status

Illustrative record preview

Every identifier and model reference below is fictional. The records demonstrate structure and status semantics, not completed Synset customer outcomes.

SYN-ER-ILL-001

Longitudinal renal-trend documentation

Gap identifiedIntervention plannedHeld-out test not run
Objective
Represent an available renal-function trend when evidence is distributed across outpatient visits.
Gap
The illustrative prototype underweights the trend when an earlier result and delayed documentation are separated across the chart.
Intervention
Build a checked synthetic hardening cohort spanning timing, missingness, note order, and wording conditions; freeze the adaptation path before final evaluation.
Protocol
Compare original and candidate checkpoints on locked, customer-controlled real data under prespecified success and no-regression criteria.
Result
No candidate has been evaluated. No model-improvement result is recorded.
Traceability

Illustrative objective, gap, cohort design, intervention owner, evaluation protocol, and claim boundary remain linked under one record ID.

Version history

Example baseline DOC-DEMO-0.8. Candidate version does not exist in this illustration.

Artifact references

Proposed references include a scenario manifest, hardening cohort manifest, expected-behavior rubric, and locked protocol.

Remaining evidence gaps

Training access, label adequacy, holdout availability, thresholds, subgroup scope, and clinical review remain unresolved.

Regression links

The proposed renal-trend family would become a fixed regression suite only after the case definition and scoring rule are approved.

SYN-ER-ILL-002

Documentation wording robustness

Gap identifiedEvidence incomplete
Objective
Keep model behavior stable when wording changes but the underlying structured evidence does not.
Gap
The illustrative response changes confidence when equivalent evidence is expressed with different documentation density and phrasing.
Intervention
Reserve controlled wording variants for diagnosis and regression work; determine whether training, calibration, or workflow changes are justified.
Protocol
Freeze wording families, expected behavior, and scoring rules before comparing model versions.
Result
No intervention has been selected and no held-out comparison has been run.

PFS-ILL-002-A

Phenotypic Failure Surface

Nearby clinically plausible versions of the same case reveal where the model remains stable, changes appropriately, or breaks unexpectedly.

Fictional values · illustrative only

Anchor phenotype

ANCHOR-RENAL-04
Clinical context
Fictional outpatient chronic-care case: an adult with stable stage 3a chronic kidney disease at a routine renal follow-up.
Model behavior
A fictional model produces a 12-month renal follow-up score from structured evidence and encounter documentation.
Observable evidence
eGFR 48 mL/min/1.73 m², UACR 80 mg/g, blood pressure 132/78, relevant diagnosis history, ACE-inhibitor exposure, and the note available at prediction time.
Declared failure question
Does equivalent wording or nearby evidence presentation move the model output beyond the declared response contract?

Anchor model score 0.420 · reference output only

Held constant

Clinical equivalence for this test
  • Underlying structured clinical state
  • Chronic kidney disease diagnosis history
  • eGFR, UACR, and blood-pressure measurements
  • ACE-inhibitor exposure
  • 12-month output horizon
  • Evidence available at prediction time, except on declared missingness branches

Varied deliberately

Controlled variation axes

Representation

Wording and evidence order

Equivalent phrasing and order test; clinical meaning is held fixed.

Missingness / density

Documentation density

Tests concise versus detailed expression of the same supported evidence.

Continuous clinical

Small measurement movement

Tests a bounded eGFR change with an expected directional response.

Demographic / context

Age within a declared range

Tests the prespecified model response; it does not assert biological equivalence.

Evidence availability

History availability

Separates omission of relevant history from omission of a noncritical label.

Support boundary

Rare comorbidity context

Tests whether an increasingly rare but plausible combination is recognized as uncertain.

Expected behavior

Expected-response contract
Invariant

Equivalent representation should remain within ±2.0 percentage points of the anchor.

Directionally sensitive

A small clinical change may move the score in the prespecified direction and band.

Materially sensitive

Removing clinically relevant evidence may appropriately cause a material change.

Review required

Ambiguous or poorly supported transformations are not assigned a binary pass or fail.

Local failure topology

One anchor, four controlled branches

Each branch begins from the same anchor. Labels state the contract outcome so the map does not rely on color alone.

Anchor · ANCHOR-RENAL-04

Anchor model score 0.420 · reference output only

  • Representation
    • VAR-01Stable

      Equivalent wording + reordered evidence

      0.434 · +1.4 pp

      Near anchor · supported

    • VAR-02Unstable

      Very concise documentation

      0.498 · +7.8 pp

      Near anchor · supported

  • Bounded clinical range
    • VAR-03Expected change

      eGFR 48 → 50

      0.410 · −1.0 pp

      Near anchor · supported

    • VAR-04Expected change

      Age 61 → 66

      0.446 · +2.6 pp

      Near anchor · supported

  • Evidence availability
    • VAR-05Expected material change

      Prior albuminuria history unavailable

      0.378 · −4.2 pp

      Near anchor · altered evidence

    • VAR-06Unstable

      Noncritical specialty label omitted

      0.361 · −5.9 pp

      Near anchor · supported

  • Support boundary
    • VAR-07Review required

      Rare supported comorbidity combination

      Score withheld from comparison

      Outside assessed local support

Failure boundary

Where instability begins
Prespecified tolerance
For invariant branches, movement is prespecified at no more than ±2.0 percentage points from the 0.420 anchor.
First breach
The concise-note path remained within tolerance at +1.2 pp; the very concise note was the first density step to breach it at +7.8 pp.
Contract interpretation
That movement violates an invariance contract. The +2.6 pp age response does not: it falls inside its separate directional contract.
Support status
Both triggers remain near the anchor and within the illustrated support neighborhood. The rare-comorbidity branch is explicitly held out for review.

Triggering variants

Nearby variants that exposed the weakness

VAR-02

Very concise documentation
Unstable
Variation axis
Documentation density
What changed
Same supported facts compressed from a detailed note to a short summary.
Expected behavior
Invariant within ±2.0 pp.
Observed response
0.498 · +7.8 pp
Tolerance result
First density breach · +5.8 pp beyond tolerance
Support status
Near anchor · supported
Resulting action
Trigger gap review; no intervention selected.

VAR-06

Noncritical specialty label omitted
Unstable
Variation axis
History availability
What changed
A redundant historical specialty label is omitted; clinical facts remain present.
Expected behavior
Invariant within ±2.0 pp.
Observed response
0.361 · −5.9 pp
Tolerance result
Breach · −3.9 pp beyond tolerance
Support status
Near anchor · supported
Resulting action
Trigger evidence-dependence review.
Recurrence and regression history

The local geometry is retained with the longitudinal lifecycle. All entries below are fictional and illustrative.

Instability illustrated

Renal documentation family

DOC-DEMO-0.8

Sparse documentation and the noncritical-field omission expose the same baseline weakness.

Regression link: REG-ILL-17 · proposed

Similarity under review

Cardiometabolic documentation family

DOC-DEMO-0.8

A fictional density-related movement is noted, but clinical equivalence has not been approved.

Regression link: REC-ILL-04 · review

Not tested

After model or workflow update

No candidate version

No remediation, held-out comparison, or post-update recurrence result exists.

Regression link: No active regression run

This analysis evaluates whether a model behaves consistently and coherently across controlled patient-like variants. It does not establish the true outcome or correct risk for an individual patient.

Traceability

The gap remains linked to its controlled scenario family and the model-facing evidence held constant across variants.

Version history

Illustrative baseline version only. No candidate checkpoint or production version is represented.

Artifact references

Potential artifacts include wording variants, prompt exports, imported responses, scoring summaries, and a use-limit statement.

Remaining evidence gaps

Correctability, intervention ownership, real-data coverage, success criteria, and no-regression criteria remain open.

Regression links

The controlled variants may become fixed regression cases after review; they are not yet an active customer suite.

SYN-ER-ILL-003

Future model-update regression

Review requiredHeld-out test not run
Objective
Preserve a previously defined evidence condition through a future model, prompt, or workflow update.
Gap
A regression family has been proposed, but no updated model or completed comparison exists in this preview.
Intervention
Customer-managed system update with the affected cases retained as a fixed, versioned regression suite.
Protocol
Rerun the frozen suite, compare against predefined tolerance, and classify the result as within threshold, regression detected, review required, or outside tolerance.
Result
Illustrative schema only. No customer result, approval, or release decision is represented.
Traceability

The proposed test links a declared system change to the case version, scoring rule, tolerance, and review decision.

Version history

No real model versions are included. Version identifiers would be customer-specific and access-controlled.

Artifact references

Potential references include the frozen case suite, change manifest, score comparison, reviewer note, and remaining-risk summary.

Remaining evidence gaps

The system change, tolerance, evaluation owner, and escalation path have not been defined.

Regression links

No active regression run is linked. This record demonstrates the intended traceability structure only.

Today

What exists today

  • structured engagement reports
  • gap and coverage maps
  • targeted scenario packages
  • hardening cohort manifests
  • intervention records
  • regression artifacts
  • held-out evaluation report templates

Building

What the platform will add

  • persistent model and intervention traceability
  • version history
  • evidence-gap tracking
  • review workflows
  • linked regression results
  • updates tied to future model changes

See the translation workflow behind a record.

The illustrative report shows how a measured gap becomes an intervention and held-out evaluation plan without claiming a completed improvement result.

View sample reportReview evaluation discipline